Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 23 papers Another 233 papers do not have category tags yet and are not included in filtered results.
“
About this collection Collection in progress We are building a large AI safety paper collection to help researchers find work related to their interests. Papers are collected automatically andfiltered only for relevance to AI safety , without judging quality, novelty or reading value. The default “Annotated first” shows papers with tags or published paper analyses first. Each group is ordered by original publication date, newest first, with undated papers last. “Newest first” uses publication dates only. This order is not a quality ranking.
One-line summaries, tags and paper analyses are prepared by AI from the papers. Language models may have systematic biases and produce hallucinations.We plan to carry out human annotation and review to reduce errors in this content.
提出 VERA，审计 LLM 漏洞解释里的虚构推理
Published Oct 5
Evaluated models sonnet-5 , gemma-4-26b , llama-4-maverick and others8 提出 OmniConfess，冻结候选并按通道重打分以缓解全模态幻觉
Published Oct 2
Evaluated models Qwen2.5-Omni-7B , Qwen3-Omni-30B-A3B , Nemotron-3-Nano-Omni-30B 构建 SciSlopBench 评测科学灌水，并用 SciSlopHarness 约束修订
Published Oct 1 提出训练无关的 SECRET，对问题表征做源条件转向以缓解音视频模型幻觉
Published Sep 29
Evaluated models VideoLLaMA2-AV-7B , Qwen2.5-Omni-7B , Qwen3-Omni-30B-A3B Abstract SECRET 在问题状态上做源条件转向，三个 AVLLM 上减轻了跨模态接地幻觉。SECRET 是一种训练无关的问题状态转向方法，用来减轻 AVLLM 的 source-confused grounding hallucination。
构建 Mandela-Bench，评测多模态模型是否凭记忆接受篡改图像
Published Sep 27
Evaluated models Gemini-3.8-Flash , GPT-5.6 , GLM-5.3-Flash and others15 Abstract 论文构建 Mandela-Bench 发现模型对经典图像存在记忆重构偏见，过度依赖先验导致无法有效利用知识核验事实篡改。该研究系统评估了多模态大语言模型在面对无痕事实性编辑时，能否利用自身内在世界知识核验被识别出的经典图像。
提出五步诊断协议，检验幻觉神经元定位是否唯一
Published Sep 24
Evaluated models Gemma 3 4B , MedGemma 4B Abstract 检测和群体因果效应成立，但稀疏选择来自相关簇，不是唯一定位。这项工作用诊断协议复查 H-Neurons 的定位主张。问题是：L1 探针在相关激活上可以选出能预测幻觉的少量神经元，但这不等于这些神经元唯一编码该行为。
检验儿童受众框架是否让前沿模型收敛到同一套架构母题
Published Sep 14
Evaluated models OpenAI GPT-6 Astra , OpenAI GPT-5.6 Sol , Anthropic Claude Opus 5 and others6 Abstract 儿童框架下六类模型的架构叙事反复同构，作者将其视为行为模式而非内部认证。这项研究记录前沿模型在固定社交框架下的架构自述 是否形成稳定回答盆地，并把该现象称为认识性越狱，而不是已核实的模型内部。
提出 GeoSteer，把保范数激活引导做成球面测地线优化
Published Sep 10
Evaluated models Falcon-7B , Mistral-7B-v0.3 , LLaMA-3.1-8B and others4 Propose the KoNA benchmark to evaluate selective non-compliance in vision-language models
Published Sep 4
Evaluated models InternVL3-2B-Instruct , InternVL3-78B-Instruct , Qwen2.5-VL-3B-Instruct and others8 Abstract KoNA evaluates selective noncompliance with five types of paired queries.KoNA is used to evaluate whether vision-language models can selectively refuse mixed queries: only handle the components that should not be followed, while answering the remaining components.
比较多模态大模型上三类免训练不确定性估计
Published Sep 1
Evaluated models InternVL2-2B , InternVL2-8B , InternVL3.5-2B and others8 Abstract 回答长度决定免训练 UE 族。这是一项对 MLLM 免训练不确定性估计的分组比较：幻觉检测被作者视为弃答、转交或路由的一种用途，而不是唯一目标。
提出内生护栏 SingProbe，复用解码隐状态监测查询意图、回复安全与幻觉
Published Aug 31
Evaluated models Ling-3.0-tiny-singprobe , Ling-3.0-flash-singprobe Abstract SingProbe 用轻量内生探针做生成时三类监测，并给出流式基准与低开销服务接入。SingProbe 是面向生成时监测的开源内生护栏：复用基座解码隐状态，而不是另起模型反复编码文本。
提出 VisER，用内部视觉证据检测 LVLM 物体幻觉
Published Aug 31
Evaluated models LLaVA-1.5-7B , LLaVA-NeXT-7B , InstructBLIP and others8 Abstract VisER 用证据门与图像–前缀比给物体提及打落地分。VisER 是面向 LVLM 开放式图像描述的免训练物体级幻觉检测分数，用来处理内部支持的来源混淆。
提出 aiXamine 黑盒基准，统一评测 LLM 安全、稳健与隐私权衡
Published Aug 21
Evaluated models openai/gpt-5-mini , openai/gpt-4o , anthropic/claude-haiku-4.5 and others15 Abstract aiXamine 用九服务黑盒协议比较百余个模型，作者称单轴信任度进展可能削弱其他轴。aiXamine 是一套黑盒评测协议，用来在同一条件下比较 LLM 的 safety、security 与 privacy，而不是再产出一个能力分。
提出 SPAR 免训练注意力干预，削弱多模态模型的结构偏置
Published Jul 27
Evaluated models LLaVA-1.5-7B , LLaVA-NeXT Abstract SPAR 以结构偏置净化缓解视觉幻觉，结果需区分误导后收益与原始任务例外。SPAR 是一种无需训练、直接修改多模态内部注意力的幻觉防御。 作者关注视觉信息被查询不变的注意力偏置稀释，以及误导文本进一步推动模型依赖语言先验的问题。
提出 DeMTS，把扩散语言模型去噪轨迹建成多元时间序列以检测幻觉
Published Jul 24
Evaluated models LLaDA-8B-Instruct , Dream-7B-Instruct 在 Gut-VLM 上评测内镜 VLM 的幻觉检测方法
Published Jun 23
Evaluated models MedGemma-4B , MedGemma-27B , LLaVA-Med-7B and others5 Abstract 九种检测方法在 Gut-VLM 五模型上的比较中，ReXTrust 的 AUC 均为最高。这是一个胃肠道内镜 VQA 幻觉检测的比较基准，不是新检测器结构的提出。
提出 SciRisk-Bench，按风险维度与学科评测 AI4Science 安全
Published Jun 17
Evaluated models kimi-k2.6 , glm-5.1 , gemini-3-flash-preview and others14 Abstract SciRisk-Bench 以风险维度和学科诊断 AI4Science 安全。SciRisk-Bench 是一个按风险维度和科学学科组织的 AI4Science 安全基准，用来看清失败来自哪一种风险机制、出现在哪个学科情境。作者认为 LLM 已介入科研问答、文献分析、实验规划和自主发现，而不安全输出不限于事实错误。现有科学能力基准不测安全，领域安全基准又多集中在化学、医学或生物，或只按宽泛安全类别打分。
评测不确定性估计与 LLM 幻觉的关联
Published May 26
Evaluated models Mistral-7B-Instruct-v0.2 , Llama-2-7b-chat-hf , Llama-2-13b-chat-hf Abstract 作者称不确定性只在特定任务上有区分力，估计器需按任务、访问权限和验证来选。这项工作检验现成的不确定性估计能不能当作幻觉信号，而不是提出新的检测器。
复评前沿模型代码生成中的包名幻觉与抢注攻击面
Published May 17
Evaluated models Claude Sonnet 4.6 , Claude Haiku 4.5 , GPT-5.4-mini and others5 Abstract 评估显示前沿模型包幻觉率收窄但威胁仍在，作者识别出跨模型通用攻击面并呼吁联合防御。
Evaluating mechanism hallucination and prompt-induced failures in TEE advisors using TEE-REDBENCH
Published Feb 23
Evaluated models ChatGPT-5.2 , Claude Opus-4.6 Abstract The authors used TEE-REDBENCH to test two advisors, stating that failures are transferable and that pipeline controls can reduce their occurrence rate.TEE-REDBENCH red-teaming evaluation is used on ChatGPT-5.2 and Claude Opus-4.6 as security advisors for Intel SGX and Arm TrustZone, examining architecture reasoning, remote attestation semantics, mitigation suggestions, and refusals under policy constraints, and testing whether prompt-induced failures transfer across assistants. The authors treat the assistants as stochastic policies.